# I Built the Ultimate Multi-Agent Workflow w/ Hermes Agent Kanban Board

Tonbi's AI Garage live demo of a durable multi-agent pipeline on Hermes Agent Kanban: scout to orchestrator rubric to parallel researchers to one Telegram human gate to builder/tester or video producer. Generalized template open-sourced.

Tags: youtube-inbox, source, hermes-agent, kanban, multi-agent, tonbi

## Summary

The hard part of multi-agent systems is coordination, not spawning agents. Hermes Kanban treats work as cards on a SQLite board (bus plus audit log). A dispatcher claims ready cards, parents block children until done, and a fleet of Hermes profiles (scouts, orchestrator, researcher, analyst, builder, tester, video producer) runs a pain-point to build-or-video pipeline with a single Telegram approval gate.

## Key points

- Problem: without a board, agents duplicate work, waste tokens, share no progress, and a crash loses state. With the board: claim, work, handoff; state lives on disk and survives restart. No agent-to-agent chat or message queues — everyone reads and writes the board.
- Card model: title, assignee (routes to a Hermes profile/model), status. One SQLite file is coordination layer, bus, and audit log.
- Dispatcher loop: ready card, claim (exactly-one), spawn assigned agent in a clean workspace, mark done, tick. Cards wait on parents in todo, auto-promote to ready; fan-out/fan-in with no glue code.
- Why it wins (speaker): durable, parallel, event-driven, self-healing (dead task reclaimed/respawned), auditable (claim/comment/completion logged).
- Workflow (content plus tools for the channel): hourly/daily scouts (X via Grok; web/Reddit/YouTube via GPT-5.5) then orchestrator dedupes and scores (bar 65/100: frequency, pain, solvable/explainable, solution gap, strategic fit) then three parallel researchers per issue (verify / context / existing solutions) then orchestrator chooses build / video / shelf then Telegram human gate (approve / shelf / modify) then builder+tester (script/skill) or video producer (slides plus speaker notes) then Telegram deliverable.
- Human gate is load-bearing: speaker would not remove it; proposals still include nonsense after all that filtering.
- Live run: gateway kept alive (tmux); forced scout sweeps; up to 18 researcher workers in parallel; about 97 tasks in one run; Telegram agent nicknamed Bulls. One self-heal: slides written to ephemeral scratch workspaces were regenerated into a persistent output dir without human prompt.
- Deliverables shown: Codex VS Code safety config verifier CLI (under 500 lines Python, report-only, redacts secrets); video outline on Claude Code subagents failing when too many MCP tools are stuffed into the prompt (allow-list custom subagents).
- Open source: generalized skeleton at Tombi Studio Hermes multi-agent workflow — research, score, action paths plus human gate; adapt with your own agent.
- Ops notes: gateway must be running for Telegram; cron 1-2 times/day not hourly; still need human polish before shipping; early Kanban versions had issues, current is fine.

## Quotes / memorable lines

- "It isn't setting up the agents themselves, it's getting them to work together on one job without stepping on each other."
- "There's no chatting between the agents. There's no message queues. Just the board that they all read and then write."
- "This is the one human gate. Everything else here happens autonomously."

## Sources

- YouTube: https://www.youtube.com/watch?v=EKVRqcpTT6s
- Raw ingest: `raw/youtube-inbox/EKVRqcpTT6s/`
- Channel: Tonbi's AI Garage https://www.youtube.com/@TonbisAIGarage

## Related pages

- [[hermes-agent]] (Kanban, profiles, cron, Telegram gateway)
- [[youtube-hermes-desktop-plugins]]
- [[perseus-vault]] / [[hermes-plugin-perseus-vault]] (shared encrypted memory across Hermes instances; cron/subagent sessions are *not* distilled into Vault)
- [[grok-bot]] (multi-bot group chat vs board-as-bus)
- [[grok-bot-vs-hermes-agent]]